Multi-Referenced Training for Dialogue Response Generation

TL;DR

This paper introduces multi-referenced training methods for dialogue response generation, utilizing diverse pseudo references and an expressive variational prior to better model one-to-many responses, leading to significant improvements.

Contribution

It proposes a novel multi-referenced training approach combining data augmentation with pseudo references and an expressive variational prior, enhancing dialogue response diversity.

Findings

Significant improvements in automated and human evaluations.

Effective use of pseudo references from pretrained models.

Enhanced modeling of one-to-many dialogue responses.

Abstract

In open-domain dialogue response generation, a dialogue context can be continued with diverse responses, and the dialogue models should capture such one-to-many relations. In this work, we first analyze the training objective of dialogue models from the view of Kullback-Leibler divergence (KLD) and show that the gap between the real world probability distribution and the single-referenced data's probability distribution prevents the model from learning the one-to-many relations efficiently. Then we explore approaches to multi-referenced training in two aspects. Data-wise, we generate diverse pseudo references from a powerful pretrained model to build multi-referenced data that provides a better approximation of the real-world distribution. Model-wise, we propose to equip variational models with an expressive prior, named linear Gaussian model (LGM). Experimental results of automated evaluation and human evaluation show that the methods yield significant improvements over baselines. We will release our code and data in https://github.com/ZHAOTING/dialog-processing.